Jaroslav emailed this issue:

On 7/9/20 9:07 AM, Jaroslav Bachorík wrote:
> Hello,
>
> Recently, after backporting JDK-8233197 to JDK8u I received a report
> from Sergey Nazarkin (cc) that the backport broke JVMTI spec basically
> emitting JVMTI_EVENT_THREAD_START events in primordial phase while the
> spec states that it should be emitted only in either start or live
> phase.
>
> But if I am reading this correctly
> (https://hg.openjdk.java.net/jdk8u/jdk8udev/hotspot/file/02b4fd2f9041/src/share/vm/prims/jvmtiEventController.cpp#l418)
> the spec is also violated there.
>
> EARLY_EVENT_BITS are including THREAD_START_BIT and it is not filtered
> out by THREAD_FILTERED_EVENT_BITS so it seems perfectly valid for
> JVMTI_EVENT_THREAD_START to appear during PRIMORDIAL phase.
>
> So I would like to get clarification before going in and start
> tinkering with the JVM initialization code order.
>
> Thanks!
>
> -JB-

I responded with the follow analysis:

It looks like that change was made by this Jigsaw change:

    JDK-8072745 Re-examine JVM TI execution phases and expectations
                on what is allowed in each phase
    https://bugs.openjdk.java.net/browse/JDK-8072745

These diffs in the bug reported jumped out at me:

@@ -1022,6 +1032,10 @@
 void JvmtiExport::post_thread_start(JavaThread *thread) {
   assert(thread->thread_state() == _thread_in_vm, "must be in vm state");
 
+ if (JvmtiEnv::get_phase() <= JVMTI_PHASE_PRIMORDIAL) {
+ return;
+ }
+
   EVT_TRIG_TRACE(JVMTI_EVENT_THREAD_START, ("JVMTI [%s] Trg Thread Start event triggered",
                       JvmtiTrace::safe_get_thread_name(thread)));
 
@@ -1050,6 +1064,10 @@
 
 
 void JvmtiExport::post_thread_end(JavaThread *thread) {
+ if (JvmtiEnv::get_phase() <= JVMTI_PHASE_PRIMORDIAL) {
+ return;
+ }
+
   EVT_TRIG_TRACE(JVMTI_EVENT_THREAD_END, ("JVMTI [%s] Trg Thread End event triggered",
                       JvmtiTrace::safe_get_thread_name(thread)));


Here's the order of phases from jvmti.h:

    /* Phases of execution */

typedef enum {
    JVMTI_PHASE_ONLOAD = 1,
    JVMTI_PHASE_PRIMORDIAL = 2,
    JVMTI_PHASE_START = 6,
    JVMTI_PHASE_LIVE = 4,
    JVMTI_PHASE_DEAD = 8
} jvmtiPhase;

However, the above diffs prevent the JVMTI_EVENT_THREAD_START
and JVMTI_EVENT_THREAD_END events from being posted earlier than
the JVMTI_PHASE_START phase... Unfortunately, I can't find the
actual changeset pushed to the Jigsaw report for JDK-8072745.


When you look at the current JvmtiExport::post_thread_start()
and JvmtiExport::post_thread_end() functions, you can see this:

    1: void JvmtiExport::post_thread_start(JavaThread *thread) {
36508: if (JvmtiEnv::get_phase() < JVMTI_PHASE_PRIMORDIAL) {
36508: return;
36508: }

    1: void JvmtiExport::post_thread_end(JavaThread *thread) {
36508: if (JvmtiEnv::get_phase() < JVMTI_PHASE_PRIMORDIAL) {
36508: return;
36508: }

$ hg log -r 36508
changeset: 36508:5f9eee6b383b
user: alanb
date: Thu Mar 17 19:04:01 2016 +0000
summary: 8142968: Module System implementation


So the Jigsaw integration pushed a slightly different change
than is shown in JDK-8072745. I don't know whether the fix
for JDK-8072745 evolved before being pushed to the Jigsaw
repos or whether another JBS issue was used to modify the
code after JDK-8072745 was pushed.

Short version: The semantics of the JVMTI_EVENT_THREAD_START
and JVMTI_EVENT_THREAD_END were changed by Jigsaw, but the
JVM/TI spec was not updated to reflect those changes.